Paths to Relation Extraction through Semantic Structure

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2614–2626 August 1–6, 2021. ©2021 Association for Computational Linguistics 2614 Paths to Relation Extraction through Semantic Structures Jonathan Yellin Omri Abend School of Computer Science and Engineering The Hebrew University of Jerusalem {jonathan.yellin,omri.abend}@mail.huji.ac.il Abstract Syntactic and semantic structure directly re- flect relations expressed by the text at hand and are thus very useful for the relation extrac- tion (RE) task. Their symbolic nature allows increased interpretability for end-users and de- velopers, which is particularly appealing in RE. Although they have been somewhat overshad- owed recently by the use of end-to-end neu- ral network models and contextualized word embeddings, we show that they may be lever- aged as input for neural networks to positive effect. We present two methods for integrat- ing broad-coverage semantic structure (specif- ically, UCCA) into supervised neural RE mod- els, demonstrating benefits over the use of exclusively syntactic integrations. The first method involves reduction of UCCA into a bilexical structure, while the second leverages a novel technique for encoding semantic DAG structures. Our approach is general and can be used for integrating a wide range of graph- based semantic structures.1 1 Introduction Early work on RE focused on pattern-based rules for capturing the structure of relation-evoking words and phrases. These rules are applied over text to identify entity relations in much the same way a regular expression would be applied to dis- cover matching text. The pattern machinery spans from simple, regular-expression like, surface pat- terns (Brin, 1999; Agichtein and Gravano, 2000), through systems that integrate both lexical features and syntactic dependencies into the pattern con- struct (Mintz et al., 2009). The PredPatt and py- BART frameworks (Zhang et al., 2017a; Tiktinsky et al., 2020) are examples of syntactic dependency based systems that leverage a set of rules defined 1Code can be found on GitHub at https: //github.com/yyellin/gcn-over-semantic- representations. over Universal Dependencies (UD; Nivre et al., 2016) to extract predicate-argument structures. In supervised RE, a multi-class classifier is trained to determine whether a relation between entities is evoked by a text. With the increased pre- dominance of end-to-end neural network architec- tures in NLP practice, it is not surprising that recent work on supervised relation extraction has focused on adapting end-to-end neural systems for the task (Peters et al., 2019; Yamada et al., 2020). However, end-to-end neural models pose interpretability and customization challenges (Conneau et al., 2018; Be- linkov et al., 2020), motivating the study of hybrid models, in which the neural architecture is fed ex- plicit structure representations. The Contextualized Graph Convolution Network (C-GCN; Zhang et al., 2018) represents a method for exposing structure representation to the machinery of a deep neural network. C-GCN uses a sentence’s UD structure as explicit input to the neural network, resulting in a model whose results are both competitive and interpretable. In this paper we explore whether we can adopt broad coverage semantic structure to the same ef- fect, and whether we are able to observe improved performance in comparison …

Content truncated to 500 words